Re: [PATCH] kunit: tool: reconfigure when the used kunitconfig changes

[David Gow <davidgow@google.com>]

On Sat, Nov 20, 2021 at 5:17 AM Daniel Latypov <dlatypov@google.com> wrote:
>
> On Thu, Nov 18, 2021 at 11:04 PM David Gow <davidgow@google.com> wrote:
> >
> > On Fri, Nov 19, 2021 at 3:03 AM 'Daniel Latypov' via KUnit Development
> > <kunit-dev@googlegroups.com> wrote:
> > >
> > > Problem: currently, if you remove something from your kunitconfig,
> > > kunit.py will not regenerate the .config file.
> > > The same thing happens if you did --kunitconfig_add=CONFIG_KASAN=y [1]
> > > and then ran again without it. Your new run will still have KASAN.
> > >
> > > The reason is that kunit.py won't regenerate the .config file if it's a
> > > superset of the kunitconfig. This speeds it up a bit for iterating.
> > >
> > > This patch adds an additional check that forces kunit.py to regenerate
> > > the .config file if the current kunitconfig doesn't match the previous
> > > one.
> > >
> > > What this means:
> > > * deleting entries from .kunitconfig works as one would expect
> > > * dropping  a --kunitconfig_add also triggers a rebuild
> > > * you can still edit .config directly to turn on new options
> > >
> > > We implement this by creating a `last_used_kunitconfig` file in the
> > > build directory (so .kunit, by default) after we generate the .config.
> > > When comparing the kconfigs, we compare python sets, so duplicates and
> > > permutations don't trip us up.
> > >
> > > The majority of this patch is adding unit tests for the existing logic
> > > and for the new case where `last_used_kunitconfig` differs.
> > >
> > > [1] https://lore.kernel.org/linux-kselftest/20211106013058.2621799-2-dlatypov@google.com/
> > >
> > > Signed-off-by: Daniel Latypov <dlatypov@google.com>
> > > ---
> >
> > Note that this patch has some prerequisites before it applies cleanly,
> > notably this series:
> > https://patchwork.kernel.org/project/linux-kselftest/list/?series=576317
> >
> > I'm also seeing a couple of issues with this, though I haven't had a
> > chance to track down the cause fully, so it could just be another
> > missing prerequisite, or me doing something silly.
> >
> > In particular:
> > - Removing items from .kunit/.kunitconfig still wasn't triggering a reconfig.
>
> This is an edge case that only comes up the absolute first time you
> switch to using kunit.py with this change.
>
> If there's no last_used_kunitconfig file, this new check doesn't do anything.
> See how it returns False when the file doesn't exist in _kconfig_changed().
>
> Given you hit the error below about last_used_kunitconfig not
> existing, I'm 99% this is what you ran into.
>
> The file is currently only generated if we actually call `make oldefconfig`.
> So if you just run `kunit.py run` a few times after this change with
> no config changes, last_used_kunitconfig won't be created, and the new
> check won't kick in.
>
> We can avoid this one-time confusion by
> * make _kconfig_changed() return True if last_used_kunitconfig doesn't
> exist, since maybe the config did change.
> * or always write last_used_kunitconfig on every invocation.
>
> The first would trigger a false positive the first time a user uses
> kunit.py after this change goes in.
> It also lightly penalizes the user for messing with `last_used_kunitconfig`.

This seems like a good compromise to me: people are likely to get this
change only after a major kernel release, and re-configuring then
(even if not strictly necessary) doesn't seem totally silly. Equally,
I think it's best for the behaviour to change exactly when the change
goes in, rather than some unspecified time afterwards.

>
> The second adds some overhead that isn't really necessary most of the time.
> It also won't help with the absolute first time you run kunit.py after
> this change.
> But it will make it so the second time onwards will have the logic enabled.
>
> So I'd personally prefer we leave it as-is.
> To most users, this will be a transparent change, so there's no
> expectations about it coming into play immediately.

As mentioned above, I'd prefer this be a little less transparent and
come into play immediately. I don't think one extra reconfigure will
be a problem for most users, and it'll be obvious it's caused by an
update. Equally, I don't expect people will mess with
`last_used_kunitconfig`, so that shouldn't be a problem?

>
> > - Running with --arch=x86_64 was giving me a "FileNotFoundError:
>
> Ah, this should be unrelated to --arch.
> os.remove() throws an exception if the argument doesn't exist.
>
> So the fix is
> + if os.path.exists(old_path)
>       os.remove(old_path)  # write_to_file appends to the file

Ah... makes sense. Let's fix this in the next revision.

> And ah, that didn't get caught by the added unit test since
> build_config() is mocked out and it's in there, no build_reconfig().
>
>

So, could we have these changes for v2:
- Reconfigure if there's no last_used_kunitconfig
- Fix the os.remove() issue if last_used_kunitconfig doesn't exist.
- Note the dependencies for this to merge cleanly in the email.

Does that sound sensible?

-- David